The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the host system using the clitg command-line tool. It instructs the agent to verify the executable and provides installation instructions using the uv package manager.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external Telegram messages and updates, creating a surface for injection. Ingestion points: Data is fetched from Telegram using commands like clitg updates watch and clitg messages list as referenced in SKILL.md and streams-and-batches.md. Boundary markers: The instructions mandate structured output formats like JSON and JSONL, but they do not define specific delimiters for separating untrusted chat content when interpolated into the AI's prompt context. Capability inventory: The skill possesses broad capabilities across its scripts, including sending messages, modifying business settings, and invoking raw MTProto methods. Sanitization: The skill implements significant procedural safeguards in writes.md, such as mandatory dry-runs, the use of confirmation tokens for critical actions, and stable idempotency keys, which help mitigate unintended executions but do not specifically filter content for injection patterns.

clitg